Bridging causal discovery and graph neural networks: a comparative perspective

Abstract

Financial markets constitute complex adaptive systems distinguished by non-linear behaviors and multifaceted, temporally evolving interconnections between assets. While traditional machine learning and statistical models often struggle to capture these dynamics due to assumptions of linearity and stationarity, Graph Neural Networks (GNNs) have emerged as a powerful paradigm for modeling the relational structure of financial ecosystems. However, existing research predominantly relies on correlation-based graph constructions, such as Pearson correlation coefficients, to define edges between assets. These approaches capture symmetric co-movements but fail to distinguish genuine directional influence from spurious associations driven by global market sentiment or common macroeconomic factors. This limitation often leads to dense, noisy graph structures that hinder the learning capability of neural networks. This thesis addresses this methodological gap by proposing a novel financial forecasting framework that integrates Causal Discovery with GNN to construct and learn from directionally meaningful graph topologies. The study employs the PCMCI (Peter-Clark Momentary Conditional Independence) algorithm to infer causal graph high-dimensional time series data. By rigorously testing conditional independencies across multiple time lags, PCMCI effectively filters out indirect effects and spurious autocorrelations, thereby identifying robust causal pathways. To capture complex market dynamics, a diverse set of features—including statistical moments, spectral analysis, and technical indicators like RSI and MACD—was engineered and standardized for the model input. The proposed framework is evaluated against traditional correlation-based baselines using a comprehensive dataset of S&P 500 constituents spanning the period from January 2020 to December 2024. To prevent look-ahead bias and ensure realistic performance assessment, the data is strictly partitioned into training (2020–2022), validation (2023), and testing (2024) sets. The model's performance is systematically assessed across both regression tasks (predicting future close prices and returns) and classification tasks (predicting directional price movement) under varying forecasting horizons (1, 7, 15, and 30 days) and connectivity thresholds. Empirical results demonstrate that the causal-based GNN framework consistently yields superior predictive accuracy, characterized by lower MSE and higher F1-scores compared to correlation-based models. This performance advantage is most pronounced at medium-term horizons; for instance, at the 30-day horizon with a high filtering threshold (0.9), the causal model achieved a significantly lower MSE compared to the correlation baseline. Topological analysis reveals that causal graphs are significantly sparser than correlation graphs—exhibiting much lower density (0.0078 vs. 0.022 at the 0.7 threshold)—effectively acting as a sophisticated noise filter. This structural sparsity mitigates the risk of "over-smoothing" in deep GNN architectures, allowing the model to propagate information along the most statistically significant pathways. Furthermore, the lower spectral radius and distinct assortativity patterns of the causal graphs suggest that the model successfully captures information flow from leading indicators to lagging assets rather than merely clustering similar stocks. This research concludes that incorporating causal reasoning as a structural prior significantly enhances the robustness, and accuracy of deep learning models in financial market forecasting.